v0.3.2 · OPEN-SOURCE PYTHON SDK

Prove what your AI agent did.

AgentLedger creates verifiable, human-readable records of consequential AI-agent actions — what changed, why it changed, who requested it, and what authority allowed it.

Try the open-source SDK See a decision record

Log → Trace → Flag Risk → Review → Approve → Export

INTEGRITY VERIFIED

decision_recordHELD FOR REVIEW

actionapprove_heloc_advanceagentunderwriter-agent@1.4.0requestersvc:originationpolicyheloc_risk@v3 · §4.2riskhigh · review requiredchangeunder_review → approved_advanceapprovalpending · action not executed

trace_9c1a…f27b · sha256:a4f9…c02b · chain intact

Observed action evidenceHuman-readable recordsAuthorization contextHash-chain verification

01 — The problem

A trace can show how the system ran. That is not the whole record.

When an agent changes an application, account, workflow, or external system, teams need more than prompts and telemetry. They need evidence of the action itself.

What a trace can tell you

Prompts and model responses
Tool-call sequence
Latency, tokens, and errors
How the system behaved

What the evidence record adds

What changed and its before/after state
Who requested and authorized it
Risk, policy, and approval status
Evidence that supports the action

02 — How it works

From agent activity to reviewable evidence.

AgentLedger connects the technical trace to the consequential action, the surrounding authority, and the review outcome.

Log

Record the consequential action and the evidence reported by the integration.

Trace

Connect the action to its agent, model, prompt, workflow, tools, and requester.

Flag risk

Preserve risk, policy, and review signals without pretending to infer intent.

Review

Capture when an action is held for human review as reported by the integration.

Approve

Record the approval, denial, or override with attribution and timestamp.

Export

Produce JSON, CSV, or Markdown records with hash-chain integrity verification.

03 — Decision record

A record an auditor can actually read.

Below: an AI underwriting assistant reviewing a HELOC application. The workflow reports elevated risk and holds the action for human review. Representative AgentLedger evidence record using fictional application data.

rec_8f3a2c1e04d9 · approve_heloc_advance

HELD FOR REVIEW✓ INTEGRITY VERIFIED

Attribution

timestamp: 2026-08-31T14:22:07Z
agent: underwriter-agent@1.4.0
model: demo-underwriting-model@v1
prompt: underwrite_heloc@v12
workflow: heloc_intake@v3.2
policy: heloc_risk@v3
requester: app_2081 · svc:origination
trace: trace_9c1a…f27b

Policy & risk

risk_level: high
policy_status: warning
approval_status: pending
action_status: held_for_review

Explanation

Applicant meets equity and credit thresholds. Debt-to-income ratio of 46.2% exceeds the heloc_risk@v3 review threshold of 43%. Per policy §4.2 as reported by the workflow, advances above this threshold are held for a licensed underwriter.

Authorization basis

policy: heloc_risk@v3 · §4.2 review_required
role: svc:origination (auto-approval disabled)

State change (proposed)

− application.status: “under_review”

+ application.status: “approved_advance” (HELD — awaiting human)

events[6] · hash sha256:a4f9…c02b · prev sha256:b21c…88de · chain intactExport JSON · Markdown · Verify

Representative AgentLedger evidence record using fictional application data. Field values are illustrative; run the real SDK to produce actual records.

04 — Positioning

Complementary to observability. Not a replacement.

Observability keeps engineering teams close to model behavior. AgentLedger keeps compliance, operations, and leadership close to the actions that matter.

Observability

Optimized for engineers debugging systems.

Prompts and responses
Latency and tokens
Errors and model performance
Technical traces
System debugging

AgentLedger

Optimized for evidence of consequential actions.

Consequential actions and decisions
Requester and authorization context
Before-and-after state
Risk, policy, and approval fields
Human-readable evidence records
Hash-chain integrity verification
Audit-oriented exports

05 — Demonstration

Watch an underwriting agent do its job — and hold when it should.

The end-to-end demo shows an AI agent receiving a request, evaluating information, recording its tool calls and decisions, identifying a policy concern, holding the action for human review, and producing a verifiable export.

Run the official demo →Read the demo script

TRACE / underwriter-agent · run_7f21c

14:22:01receive_task · app_2081

14:22:02log_tool_call · verify_income · 312ms

14:22:03log_tool_call · credit_report · returned

14:22:04log_tool_call · underwriting_metrics · dti 0.462

14:22:05log_decision · high risk · DTI_OVER_THRESHOLD

14:22:06log_action · held_for_review · rec_8f3a…04d9

14:22:07✓ complete_trace · chain intact

06 — Developers

Generate your first verified record in 60 seconds.

AgentLedger is an open-source Python SDK. Clone the repo, run the end-to-end demo, and inspect the generated JSON, Markdown, and CSV evidence exports.

Try the open-source SDK →Documentation

v0.3.2 · Python 3.9+ · Developer preview

quickstart.py

# Clone and run the official demo
# git clone https://github.com/Camleader/agentledger-v01.git
# python -m pip install -e ".[dev]"

from agentledger import AgentLedger

ledger = AgentLedger(
    storage_path="events.jsonl",
    trace_storage_path="traces.jsonl",
)

trace = ledger.create_trace(
    workflow="heloc_intake",
    agent_name="underwriter-agent",
    entity_id="app_2081",
)

ledger.log_tool_call(
    agent_name="underwriter-agent",
    tool_name="underwriting_metrics",
    input_data={"app_id": "app_2081"},
    output_data={"dti": 0.462},
    trace_id=trace["trace_id"],
)

ledger.log_decision(
    agent_name="underwriter-agent",
    output_data={"decision": "hold"},
    reason_codes=["DTI_OVER_THRESHOLD"],
    trace_id=trace["trace_id"],
    risk_level="high",
    review_required=True,
    policy_status="warning",
    approval_status="pending",
)

ledger.complete_trace(
    trace_id=trace["trace_id"],
    outcome="held_for_review",
    approval_status="pending",
)

ledger.export_markdown_report("audit_report.md")
integrity = ledger.verify_hash_chain()  # ✓ chain intact

07 — Design partners

Shape the evidence layer with us.

We’re inviting a small number of software companies deploying AI agents into regulated, high-trust, or operationally consequential workflows to become early design partners. Design partners help define the evidence their auditors and reviewers actually need.

Identify consequential agent actions

Instrument AgentLedger

Define required evidence fields

Capture authorization & attribution

Generate human-readable records

Provide feedback that shapes the roadmap

08 — Founder

Cameron Roberts

Founder, AgentLedger

cameron@agentledgerai.com

Agents are being handed more responsibility. The record of what they do should keep up.

AgentLedger exists to give organizations a trustworthy history of AI-agent actions — one that a compliance officer, an auditor, and an engineer can all read and agree on.

It’s early. It’s built in the open. If your team is deploying agents into consequential workflows and you care about being able to prove what they did, I’d like to talk.